Case study: how a French diesel-parts store put AI in front of 439 support conversations a month

Injecteur France sells remanufactured diesel injectors to workshops and drivers across France. Support questions are repetitive and high-stakes at the same time: does this part fit my engine code, where is my order, how does the core deposit work — and when the answer is wrong, a workshop installs the wrong injector.

Here is what one month of production looks like with Zedinga AI as the first line of support. These are real numbers from the last 30 days, not projections.

The numbers (last 30 days)

• 439 customer conversations handled by the AI first — on the website widget, in French.
• 237 leads captured — contact details collected in chat and landed in a pipeline the team actually works.
• 320 conversations routed to the human team — by design: the bot's conversation flow sends concrete purchase and quote requests straight to the team's Telegram group instead of letting the AI guess prices.
• 12 unanswered questions — under 3% of conversations produced a question the AI could not ground in the knowledge base. Each one lands in a review queue and becomes training material.

What the bot actually does

Compatibility without guessing. The system prompt enforces a hard rule: injectors are matched by the part number stamped on the body, never by car model alone. If the visitor only names a vehicle, the bot asks for the reference — and answers from a catalog of the store's real products, synced from Shopify.

Order status in chat. A customer asks "où est ma commande ?", verifies with their order number and email, and gets the live status with the carrier tracking link — pulled from Shopify at that moment, shown only after the order-bound verification passes.

Order actions, not just answers. The store runs Zedinga's AI Actions: a verified customer can cancel an unshipped order (with automatic refund), change a delivery address, or open a return request — each one confirmed with a 6-digit code sent to the email on the order, capped at a merchant-set amount, and logged in an audit trail. Orders above the cap, or orders a guardrail blocks (already shipped, already refunded), are handed to the team with an honest explanation.

Honest handoff. When a conversation needs a human — a price negotiation, a complex diagnosis — the flow hands it over and says so. The team answers from the dashboard or the mobile app; nothing pretends to be resolved when it isn't.

Why the unanswered rate stays low

The bot runs on strict grounding: if retrieval can't support an answer, it says it doesn't know and offers the team — it does not improvise. That discipline is what keeps the unanswered queue at a dozen questions a month instead of a flood of confident wrong answers. Combined with a curated knowledge base (the store's pages, policies and a part-number catalog), the result is a support surface the owner doesn't have to babysit.

The setup

Everything above is configuration, not custom code: a French-language widget matched to the store's brand, a conversation flow for quote requests, Shopify connected for catalog + orders + actions, strict grounding on, and notification routing to Telegram. The same setup is repeatable for any parts store — the knowledge base and guardrail prompt are the work, and they are done once.

---

Numbers are from the store's live dashboard for the 30 days ending 2026-10-20, shared with the store's permission. If you run a parts store and want this setup, talk to us — the first conversation is free.